Equity Research Decision Intelligence
Design a research platform that turns a point-in-time document corpus into auditable next-day equity decisions.
Problem statement
For this exercise, assume the source-backed system identity and the explicitly bounded practice contract that follow. Design an end-to-end equity research decision platform.
It ingests roughly ten years and four million research reports, news items, and earnings materials; reconstructs the information known about each company at a requested point in time; and produces an auditable next-day trading decision.
The operating team is one research engineer and one platform engineer, with eight B200-class accelerators available. Keep the design feasible for that team, distinguish research-time experimentation from the production decision path, and explicitly decide where open-source models or hosted APIs are appropriate.
What to build
- Ingest, normalize, and version research reports, news, and earnings materials without losing event time or publication time.
- Reconstruct a company view using only information that was available by the requested cutoff.
- Produce a next-day decision with the evidence, model version, features, and configuration needed to reproduce it.
- Let researchers evaluate candidate retrieval and inference changes without silently altering production history.
At scaleExercise boundary
- The historical corpus spans about ten years and four million source documents.
- Eight B200-class accelerators are available, so scheduling and isolation must prevent exploratory jobs from starving daily production work.
- The system is operated by one research engineer and one platform engineer, favoring a small number of observable managed components.
- A daily decision must use a point-in-time snapshot and finish before the next trading session.